Hybrid model navigation method for autonomous underwater vehicle

Accurate navigation and localization are essential for Autonomous Underwater Vehicles (AUVs). However, the unknown modeling errors and nonlinear errors will affect the AUV positioning accuracy. Meanwhile, the marine environment changes may not be accurately sensed by AUV. Therefore, this paper proposes a Hybrid Model (HM) navigation methodology for AUV to reduce the impact of unknown errors and better predict the future states simultaneously. Firstly, an error correction sub-model based on Sequence to Sequence (Seq2Seq) predicts AUV pseudo displacements. The pseudo displacements are augmented to the observation vector to correct the unknown errors in the state estimation. Secondly, a state regression sub-model based on Gaussian Process Regression (GPR) is utilized to capture the motion trends from the historical data and regress the state variation, which is used to resist the unknown disturbances in the changeable marine environment. The sub-models work parallel with the master model, benefiting from the Interacting Multiple Model (IMM). We compare the performance of the proposed HM to Extended Kalman Filter (EKF), Unscented Kalman Filter (UKF), and IMM-EKF by using the actual experimental data of Sailfish 210 AUV. The experimental results show that the proposed HM algorithm achieves superior navigation accuracy and good fault tolerance capability.